Mistral Medium 3.5: The Dense AI Redefining Chat and Coding

Mistral, the French company specializing in artificial intelligence, recently unveiled its latest model, Medium 3.5. This dense model, featuring 128 billion parameters, merges chat, reasoning, and coding capabilities into a single entity. It offers a context window of 256,000 tokens, simplifying its deployment despite a high inference cost.

Mistral's development tool, Vibe, also benefits from a significant update. It now integrates asynchronous cloud agents capable of handling routine tasks such as bug fixes, module refactoring, test generation, and dependency updates without requiring constant supervision. These agents operate in isolated environments and can connect to popular services like GitHub, Slack, Linear, Jira, Sentry, and Teams. Additionally, local sessions can be transferred to the cloud, along with their history, task status, and approvals, and each agent can open a pull request once completed. The cloud version of Vibe is built on workflows derived from Mistral Studio, which the company initially developed in-house and for enterprise clients.

Meanwhile, Mistral's AI assistant, Le Chat, introduces a "work mode" designed for multi-step workflows. This mode automatically activates connectors to emails and calendars while requiring explicit user approval for sensitive actions such as sending a message or writing to external systems. The work mode is available on the Pro, Team, and Enterprise plans.

Mistral has chosen to distribute Medium 3.5 under a "Modified MIT License," replacing the Apache 2.0 license it previously used. This new license allows for both commercial and non-commercial use, but with certain restrictions for high-revenue companies. The model weights are available on Hugging Face, and access via the API is priced at $1.50 per million input tokens and $7.50 per million output tokens.

The Medium 3.5 model follows the industry's trend of integrating reasoning as a parameter for each query, with a "reasoning_effort" parameter allowing users to switch between quick responses and a heavier mode for complex tasks. Mistral has also reworked the visual encoder from the ground up to better handle variable image sizes and aspect ratios.

In Mistral's internal benchmarks, Medium 3.5 achieved scores of 77.6% on SWE-Bench Verified and 91.4% on T3-Telecom. The model replaces Medium 3.1 and the reasoning model Magistral in Le Chat, as well as Devstral 2 in the Vibe CLI.

Mistral is aware that there are less costly approaches. For example, Mistral Large 3 uses a Mixture of Experts (MoE) configuration with a total of 675 billion parameters but only activates 41 billion per token. Mistral Small 4 has 119 billion parameters and activates only 6 billion. Competitors like Deepseek and Qwen have long directed their models toward MoE, as it allows for cheaper inference while maintaining similar quality.

Finally, Mistral indicates that the model can be self-hosted on four GPUs, although this is likely out of reach for most users outside of well-equipped data centers.